The Drug-Target interaction (DTI) prediction is a crucial step in drug discovery and repositioning. Although traditional biochemical experiments yield highly reliable results, their high cost and low efficiency limit large-scale screening. Most existing computational methods rely on multi-dimensional features to achieve good predictive performance but overlook in-depth exploration of the intrinsic features of drugs and targets. This study introduces SSTDTI, a novel approach that computes similarity coefficients by analyzing substructure features of drugs and the functional annotations of targets. These coefficients are utilized to construct directed, weighted adjacency graphs for both drugs and targets. The model employs convolutional operations to efficiently extract features from the resulting graph structures. Additionally, our newly designed attention mechanism facilitates mutual layer-wise fusion encoding between drugs and targets, enabling the extraction of deeper interaction features. By integrating multi-scale features, SSTDTI outperforms existing methods across multiple benchmark datasets, with ablation studies further validating the efficacy of its constituent modules.